MiniMax M2.5

MiniMax M2.5 was announced on 12 February 2026 and is the strongest model the company built on the M2 body. The architecture is untouched — 62 layers, 256 experts, 8 routed per token — and the parameter count is 228,703,644,928, which is 13,880,064 more than M2, M2.1 and M2.7, all three of which are identical to the byte. That fourteen-million-parameter difference is the only structural change anywhere in the line. The results moved a long way. MiniMax reports 80.2% on SWE-bench Verified, 51.3% on Multi-SWE-bench and 76.3% on BrowseComp with context management, and says the model finished the SWE-bench Verified run 37% faster than M2.1. The training was reinforcement learning across more than 200,000 real environments in over ten programming languages, and the company describes a behaviour that emerged rather than being designed: before writing code, the model writes a specification, decomposing structure and interface the way an architect would. The claim MiniMax led with was economic, not technical — that M2.5 is the first frontier model where cost stops mattering, at roughly one dollar to run continuously for an hour at 100 tokens per second, or thirty cents at half that rate. The list price is $0.30 per million input tokens and $1.20 per million output, unchanged from M2 sixteen months earlier. What did change was the paperwork. M2 and M2.1 were MIT. M2.5 shipped under the MiniMax Model License, a bespoke agreement dated 13 February 2026 that permits redistribution but requires a specific attribution notice to travel with every copy. It was the first step away from standard open-source terms in this family, and the next model went considerably further.

▸Videos

No videos yet.

▸Reviews